7th International Conference on Spoken Language Processing
September 16-20, 2002
In this paper, we propose a new approach to online adaptation of continuous density hidden Markov model (CDHMM) based on speaker space model evolution. The speaker space model which characterizes the a priori knowledge of the training speakers is effectively described in terms of the latent variable model such as the factor analysis (FA) or probabilistic principal component analysis (PPCA). The latent variable models are employed to provide not only the speaker space model but also a joint prior distribution of CDHMM parameters, which can be directly applied to the maximum a posteriori (MAP) adaptation framework. We establish an online adaptation scheme based on the quasi-Bayes (QB) estimation technique which incrementally updates the hyperparameters of the speaker space model and the CDHMM parameters simultaneously. In a series of speaker adaptation experiments on the task of continuous digit recognition, we demonstrate that the proposed approach not only achieves a good performance for a small amount of adaptation data but also maintains a good asymptotic convergence property as the data size increases.
Bibliographic reference. Kim, Dong Kook / Kim, Nam Soo (2002): "Markov models based on speaker space model evolution", In ICSLP-2002, 1393-1396.